Hi Guys,
Performance meeting with be at 8:00am PST as usual. Not much on the
agenda for today. A bug was discovered in the way that we export kernel
rbd disk geometry that results in mkfs.xfs chosing large sunit and
swidth values, ultimately resulting in a large (17!) number of AGs and
high random IO behavior on RBD images. Manually decreasing the AG count
at filesystem creation time results in a performance warning from
mkfs.xfs, but actually improves small file creation performance by up to
2X in our tests. Nothing else on the agenda for today (yet!).
Please see the etherpad for meeting details:
https://pad.ceph.com/p/performance_weekly
Thanks,
Mark
